fix: knowledge retrieval accuracy and remaining brand/PoC leaks in procedure docs

- Fix the demo knowledge provider's tokenizer: a plain [a-z0-9]+ regex silently
  dropped accented characters, splitting French words like "véhicule" into "v" +
  "hicule" and mangling retrieval for nearly every French query. Now matches the
  Latin-1 accented range too.
- Reweight section scoring so the body match (the actual substance of a section)
  outranks a heading/title match (a shallow structural hint) rather than the reverse
  -- confirmed via the brief's exact validation question that the old weighting
  misranked the damage procedure behind a topically-adjacent document in all three
  languages (nl-BE: a checkout section; en-GB/fr-BE: the return procedure), purely
  because a generic word like "vehicle"/"voertuig" happened to sit in a heading/title.
- Remove leftover "MobilityOps" and "PoC" mentions from 5 English and 4 NL/FR
  procedure documents -- knowledge-base prose is visible UI content and was missed by
  the earlier rebrand.
- Add regression tests: the brief's exact NL/EN/FR damage question must ground on the
  damage procedure as the *primary* source (not just appear in the top 3), and no
  procedure file may contain "MobilityOps" or "PoC".

151 backend tests, Ruff, mypy green.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

backend/app/services/knowledge/demo.py

@@ -35,7 +35,11 @@ STOPWORDS_BY_LANGUAGE: dict[str, set[str]] = {
     },
 }
 
-_WORD_RE = re.compile(r"[a-z0-9]+")
+# Includes the Latin-1 accented-letter range (à-ö, ø-ÿ) so French/Dutch words with
+# diacritics (véhicule, réservation, geëscaleerd) tokenize as one word instead of
+# splitting apart at the accented character -- a plain [a-z0-9]+ pattern silently
+# drops every accent and fragments the word either side of it.
+_WORD_RE = re.compile(r"[a-zà-öø-ÿ0-9]+")
 
 
 def _stem(word: str) -> str:
@@ -218,17 +222,28 @@ class DemoKnowledgeProvider:
     def _score(
         self, query_tokens: set[str], section: ScoredSection, idf: dict[str, float]
     ) -> float:
+        # The section body is the strongest relevance signal -- it's the actual
+        # substance a heading or title can only hint at -- so a body match is weighted
+        # *above* heading/title matches, not below them. The previous 3x/2x/1x
+        # (heading/title/body) ordering let a single generic word in a heading (e.g.
+        # "vehicle", present in nearly every section) or a document's own title
+        # outrank a section whose body genuinely covers multiple, more distinctive
+        # query terms -- confirmed to misrank the brief's exact validation question in
+        # every one of the three languages (see docs/fleet-ops-correction/
+        # current-gap-audit.md and i18n-inventory.md): nl-BE picked a checkout section
+        # over the damage procedure, en-GB and fr-BE picked the return procedure over
+        # the damage procedure, purely from heading/title overlap on common words.
         score = 0.0
         for token in query_tokens:
             token_idf = idf.get(token, 0.0)
             if token_idf == 0.0:
                 continue
-            if token in section.heading_tokens:
+            if token in section.body_tokens:
                 score += 3 * token_idf
-            elif token in section.document.title_tokens:
+            elif token in section.heading_tokens:
                 score += 2 * token_idf
-            elif token in section.body_tokens:
-                score += token_idf
+            elif token in section.document.title_tokens:
+                score += 1.5 * token_idf
         return score
 
     def ask(

backend/tests/test_knowledge.py

@@ -1,11 +1,51 @@
 from __future__ import annotations
 
+import re
+from pathlib import Path
+
 import httpx
 
+from app.core.config import get_settings
 from app.services.knowledge.demo import DemoKnowledgeProvider
 from app.services.knowledge.ragcore import RAGcoreKnowledgeProvider
 
 
+def test_brief_exact_damage_question_in_all_three_languages():
+    # The exact validation questions from docs/fleet-ops-correction/current-gap-audit.md
+    # -- each must ground on the damage procedure as its *primary* (top-ranked) source,
+    # not merely appear somewhere in the top-3, and the source/version/section/excerpt
+    # must all come from that same-language document (never an English fallback).
+    provider = DemoKnowledgeProvider()
+    cases = {
+        "nl-BE": "Wat moet ik doen wanneer een voertuig beschadigd terugkomt?",
+        "en-GB": "What should I do when a vehicle returns with damage?",
+        "fr-BE": "Que dois-je faire lorsqu'un véhicule revient endommagé ?",
+    }
+    for language, question in cases.items():
+        answer = provider.ask(question, f"test-brief-{language}", language)
+        assert answer.evidence_state == "grounded", language
+        assert answer.sources, language
+        assert answer.sources[0].document_id == "damage-procedure", (
+            f"{language}: expected the damage procedure as the primary source, "
+            f"got {answer.sources[0].document_id!r}"
+        )
+        assert answer.answer
+        assert answer.sources[0].excerpt
+
+
+def test_knowledge_procedures_never_mention_mobilityops_or_poc():
+    # Section 2 of docs/fleet-ops-correction/current-gap-audit.md: the visible brand
+    # name is exactly "Fleet Ops", and "PoC" must never appear in visible content --
+    # including the demo knowledge base, not just the frontend.
+    procedures_dir = Path(get_settings().knowledge_dir)
+    offenders = []
+    for path in sorted(procedures_dir.glob("*/*.md")):
+        text = path.read_text(encoding="utf-8")
+        if "MobilityOps" in text or re.search(r"\bPoC\b", text):
+            offenders.append(str(path))
+    assert offenders == []
+
+
 def test_s6_damage_question_is_grounded_with_expected_sources():
     provider = DemoKnowledgeProvider()
     answer = provider.ask(